/newsroom/taxonomy/term/10219/all en /newsroom/channels/news/science-artificial-intelligence-faces-reproducibility-crisis-285180 <p>At the AAAI meeting, <strong>Peter Henderson</strong>, a computer scientist at ƻ����Ժ in Montreal, showed that the performance of AIs designed to learn by trial and error is highly sensitive not only to the exact code used, but also to the random numbers generated to kick off training, and to “hyperparameters”—settings that are not core to the algorithm but that affect how quickly it learns.<br /> <a href="http://science.sciencemag.org/content/359/6377/725.summary">Science</a></p> Mon, 19 Feb 2018 18:08:44 +0000 nathan.menezes@mail.mcgill.ca 32975 at /newsroom